CS491 Great Principles Assignment A8 Due 11/4/02 For the next time, we'll talk about the principles of performance measurement and prediction of computer systems and networks. Everyone is interested in how long it will take to run a computation on a given configuration of hardware and in the presence of contention from other jobs. In preparation, read the queueing network tutorial linked to the CS471 web resources section (it is a reprint of a pair of American Scientist essays I wrote) and also the queueing module of the ArtOS book (cne.gmu.edu/pjd/ArtOS). Write short answers to these questions: (1) Why is it a concern to know how long it will take a system to respond to a user's request? And how many user jobs are completed every hour? (2) Flow balance is an assumption used in most performance analyses. What is flow balance? How much error can be introduced by assuming flow balance for a model of a system that is not flow balanced? (3) Under what conditions is it true that the bottleneck of a system is the server with the largest demand? How much error would be introduced by assuming this for a real system? (4) In the MVA algorithm, the first equation says that response time at a service can be computed as the service time multipled by a stretch factor equal to 1 plus the queue length observed by the new arrival to the server. That queue length is computed as the overall average queue length when there is one less user in the system. The assumption that the arriving user sees the same average size queue as the outside observer with one less user in the system is called homogeneity. Homogeneity asserts that the arriving user acts as an outside observer in a system where that user is excluded from the system population. Homogeneity is sometimes phrased as "the off line behavior of a server is the same as its on-line behavior." Comment on how this latter definition equates to the former. What kinds of errors might be introduced by assuming that the servers are homogeneous in this way? If the real system really is homogeneous, would the MVA equations compute the measured values of throughput and response time?